(560f) Machine Learning-Driven Vaccine Design Against Highly Mutable Pathogens

Vaccines play a critical role in both preventing and eliminating global pandemics. This is especially true in the case of highly mutable pathogens like HIV and influenza, which present unique challenges to vaccine design. One promising strategy for designing vaccines against highly mutable pathogens is to target regions on the surfaces of the pathogenic proteins that are functionally important (e.g., receptor binding sites for adhering to/entering host cells), and thus cannot be so easily mutated by the pathogen. Antibodies that can bind to these ‘conserved’ regions on pathogenic proteins are called broadly-neutralizing antibodies, or bnAbs. BnAbs have now been isolated from people naturally infected with many different highly mutable pathogens. However, in most cases it remains unclear how to elicit these same antibodies by vaccination, in part because of the complexity of the key biological processes that occur but also because of the vast design space for rationally designing such vaccines. Design variables include the number of immunizations to administer, the number of Ags (pathogen-like proteins) to administer in each immunization, the concentration of each Ag in each immunization, and the sequences of the Ags. This talk will discuss a novel computational/data science-driven approach to efficiently traverse this vast design space. Our approach features agent-based modeling of cell-cell interactions combined with deep reinforcement learning (DRL) to steer the evolution of antibodies into bnAbs following vaccination. Using a rewards and punishments-based system, results show that DRL is able to rapidly identify optimal vaccination protocols that maximize the formation of bnAbs against highly mutable pathogens.